c_prediction_snn.m

神经网络的工具箱, 神经网络的工具箱,

function [c_point_pred, noise_net, noise_record] = c_prediction_snn(elevel, nets, datasets, alpha, one_net, blocksize)
%C_PREDICTION_SNN Estimate c_prediction
%
%  Syntax
%
%   [c_point_pred, noise_net, tr_info_noise_net] = ...
%                c_prediction_snn(e_level, nets, datasets, alpha)
%
%  Description
%
%   C_PREDICTION_SNN takes
%     e_level   - error level.
%     nets      - [1 x M] net_structs of trained networks with cost 
%                 function WCF_SNN.
%     datasets  - [1 x M] matrix of dataset_structs containing
%                 information on the data used in training 'nets'. 
%     alpha     - [1 x M] matrix of network weighting factors.
%   and returns
%     c_point_pred   - c_prediction
%     noise_net      - a net_struct for estimating noise on input data.
%     tr_info_noise_net - information about training of 'noise_net'.
%
%  See also
%
%   PREDICTION_SNN, C_CONFIDENCE_SNN, BALANCE_SNN
%

data = datasets(1).data;
M = size(nets,2);
N = size(nets(1).biases{nets(1).numLayers},1);
[NI, MU] = size(data.P);
g = getg_snn(nets(1), data);
errf = nets(1).costFcn.fn;
if nofieldorempty_snn(data, 'useT')
   data.useT = ones(N, MU);
end

% compute conf_err_estimate(x^nu)
[y_av, conf_err_estimate] = simff_avr_snn(nets, alpha, data.P);

% compute pred_var_estimate(x^nu) and noise_net

% compute average net
if (nargin < 5)
   stdout_snn('Summarizing networks to one for error estimation...\n');
   one_net = summarize_snn(nets, data, alpha);
   stdout_snn('...done\n');
end

if (nargin <6)
   blocksize = 1;
end

y_validation = unbiased_average_snn(nets, alpha, datasets);
noise_data.P = data.P;
%#function se_snn
%#function relerr_snn
%#function loglikelihood_snn 
%#function crosslogistic_snn
%#function crossentropy_snn
if isstr(errf)
   %noise_data.T = max(feval(errf, data.T, y_validation) -conf_var_estimate, 0);
   noise_data.T = feval(errf, data.T, y_validation);
elseif iscell(errf)
   for i = 1:N
      %noise_data.T(i,:) ...
      %    = max( feval(errf(i), data.T(i,:), y_validation(i,:)) ... 
      %            - conf_var_estimate(i,:), 0);
      noise_data.T(i,:) = feval(errf{i}, data.T(i,:), y_validation(i,:)); 
   end
else
   error('Incorrect format for output errorfunction');
end
noise_data.T(find(noise_data.T == 0)) = eps;

if isfield(data, 'useT')
   noise_data.useT = data.useT;
end
if isfield(data, 'gmu')
   noise_data.gmu = data.gmu;
end
stdout_snn('Computing NN for error estimation...\n');
[noise_net, noise_record] = train_noise_snn(one_net, noise_data, blocksize);
stdout_snn('...done\n');
pred_err_estimate = simff_snn(noise_net, data);

% compute total_err_estimate(x^nu)
total_err_estimate = max(conf_err_estimate, pred_err_estimate);

% compute total_err(x^nu)
total_err = zeros(N, MU);
if isstr(errf)
   ind = find(data.useT);
   total_err(ind) = feval(errf, y_validation(ind), data.T(ind));
elseif iscell(errf)
   for i = 1:N
       mu = find(data.useT(i,:));
       total_err(i,mu) = feval(errf{i}, y_validation(i,mu), data.T(i,mu));
   end
else
   error('Incorrect format for output errorfunction');
end

% compute c_point_pred
for i = 1:N
    mu = find(data.useT(i,:));
    d3 = g(i,mu);
    d3 = d3(:)./(sum(d3(:)));
    [d1, d2] = ...
    sort(-(total_err(i,mu)./(total_err_estimate(i,mu))));
    c_point_pred(i,1) = -d1(min(find( cumsum(d3(d2))>elevel ))); 
end